PaperlayerView exact revision on arXiv

Source-linked AI summary

Sensorless damage-safe grasping

Yusei Shuto, Danilo Vasconcellos Vargas

arXiv:2608.23983v1cs.RO

TL;DR

Fruit harvesting needs secure grasps without bruising produce whose stiffness varies widely, while dedicated tactile sensing adds deployment complexity. This paper uses encoder and motor-effort signals to conservatively bound compression strain, achieving at least 98% grasp at 0% damage across the certified strain range, unlike fixed-force baselines.

  • Problem

    Fruit stiffness varies widely within species, so fixed-force grasping cannot simultaneously secure and preserve produce without dedicated adaptive sensing.

  • Method

    A sensorless controller estimates compression strain from encoder position, motor effort, and a lower stiffness bound, stopping at a user-specified limit ε.

  • Results

    ≥98% grasp at 0% damage held across the entire ε∈[0.7, 2.0]% range at k=4000 N/m, which neither fixed-force baseline reached.

  • Takeaways & Limitations

    The approach provides a certified, operator-interpretable damage bound while making closing speed an explicit throughput–gentleness control.

  • Takeaways & Limitations

    Rigid-pad grasping remains limited in the very-soft regime, where soft-pneumatic or compliant grippers remain standard.

Abstract

from arXiv · show

Robotic fruit harvesting must hold produce securely without bruising it, yet compression stiffness varies several-fold with ripeness within a single species, so no fixed grip force spans the range. Rather than tune force, we bound deformation: a controller closes the gripper until the object's estimated compression strain reaches a user-specified limit $\varepsilon$, using only the encoder position and motor-effort signal on every servo gripper---no tactile or force-torque sensor. Dividing an effort-based contact force by a lower bound on object stiffness makes the stop provably conservative---true compression stays at or below $\varepsilon$---for any $\varepsilon$ above a contact-detection strain floor we identify and quantify: robust detection itself spends compression, linearly in closing speed, making speed an explicit throughput--gentleness knob. Unlike a hand-tuned force threshold, $\varepsilon$ is a certified, size-scaling, operator-interpretable damage limit, and a ready safe-action parameter for learned grasping policies. In MuJoCo simulation over a realistic fruit-stiffness range, under a sensor-noise model calibrated to the real servo, the controller holds $\ge 98\,\%$ grasp at $0\,\%$ damage across all medium-to-firm stiffnesses for the entire certified $\varepsilon$ range, which neither fixed-force baseline attains; on stiffness-graded 3D-printed TPU cubes it matches baseline grasp success at roughly half the grip force and cuts soft-object damage from $100\,\%$ to $40\,\%$.

1. Introduction

This work addresses the conflict between secure fruit grasping and damage avoidance by regulating compression strain with encoder and motor-effort signals alone. It provides a conservative strain guarantee above a quantified detection floor and validates lower damage than fixed-force grasping across varying stiffnesses.

  • Problem: Fruit harvesting must balance holding produce securely during transport against avoiding damage, while tomato stiffness varies from roughly 2000 N/m to 10 000 N/m within one species.A force secure for firm fruit can crush overripe fruit.
  • Motivation: Dedicated tactile or vision–tactile sensing adds cost, fragility, and integration complexity, motivating adaptation from encoder and motor-effort data already available on actuated grippers.The approach uses no tactile or force-torque sensor.
  • Approach: The controller regulates grasp compression to a user-specified strain bound ε using only proprioceptive signals, rather than adapting grip force to each fruit.Compression strain is estimated from an effort-based force proxy and a known lower stiffness bound.
  • Guarantee: True compression strain provably stays at or below ε for any ε above the contact-detection floor, using only a lower bound on object stiffness and no per-object calibration.The guarantee is conservative across the fruit-stiffness range.
  • Detection limit: Noise-robust contact detection itself spends compression linearly with closing speed, making speed an explicit throughput–gentleness knob and bounding the attainable ε from below.The quantified floor also explains the hardware observations.
  • Validation: Simulation and hardware validation across varying object stiffnesses demonstrates lower damage than fixed-force grasping at comparable grasp success, using simulation noise calibrated to the hardware.The introduction positions this as validation of the deformation-bounded approach.

2. Related Work

Related work establishes that fruit harvesting must treat grasping as a damage-sensitive task, while existing deformable-object methods often rely on tactile hardware or categorical grasp labels. Proprioceptive force estimation from encoder and motor-effort signals therefore provides the relevant sensorless alternative.

  • Damage-sensitive harvesting: Fruit harvesting research emphasizes that grasping is critical because fragile produce bruises easily and damage sharply reduces market value.Harvesting pipelines typically detect and localize fruit before driving an end effector to detach it.
  • Tactile grasp assessment: Han et al. combine GelSight tactile sensing with RGB vision to classify deformable-object grasps as safe, slippery, or damaging, rather than exposing a user compression-strain parameter.Their reported success varies across object shapes, including 38 % on bananas and 90 % on round fruit.
  • Tactile grasp assessment: 38 % on bananas and 90 % on round fruit are the reported success rates for Han et al.’s comparable deformable-object grasping method.Similar visual-tactile fusion methods also assess grasp state on deformable objects, but rely on specialized tactile hardware.
  • Proprioceptive force estimation: Actuator-based interaction-force estimation is established through generalized momentum and motor current, with later refinements using disturbance observers, learned friction models, and deep temporal networks.This principle has also been applied to grasping with underactuated grippers.

3. Method

The method estimates compression strain from encoder position and motor effort, then proportionally closes or holds the gripper until a user-specified strain bound is reached. Dividing the effort-based force proxy by a stiffness lower bound makes the stop conservative without tactile or force-torque sensing.

  • Sensorless strain control: The controller uses only 50-Hz jaw position and motor-effort signals to estimate compression strain and command normalized jaw velocity, without tactile or force-torque sensing.Motor effort is represented by current in simulation and a load-register duty-cycle proxy on the Feetech servo.
  • Sensorless strain control: The method detects contact by effort-amplitude threshold, estimates strain by dividing an effort-based force proxy by k_minD, and drives that estimate toward ε.The pipeline uses fruit diameter D and stiffness lower bound k_min as external inputs.
  • Control law: The same proportional law closes or holds the jaw during grasping and re-closes during lift if estimated strain falls because of arm-joint yield.With v_release = 0, the experiments use a close-or-hold controller whose action clips to zero after brief overshoot.
  • Operating parameters: The operator specifies ε, object diameter D, and k_min as task or object-class inputs, while closing speed v controls the throughput–gentleness trade-off.No per-fruit stiffness measurement is required, and the controller tolerates effort-coefficient error within the stiffness margin.
  • Stiffness estimation: The online Hookean stiffness fit is disabled because damage-free stops provide insufficient displacement for reliable identification, so reported results use the pure-k_min controller.The fit failed its acceptance gate without noise and sometimes produced several-fold errors under sensor noise.
  • Conservative safety bound: A lower bound k_min set to the softest produce makes estimated strain over-bound true strain, so stopping at ε certifies true compression at or below ε.The guarantee applies for ε above the contact-detection floor; transient overshoot is clipped rather than corrected.

4. Experiments

Experiments combine MuJoCo simulation across fruit sizes and stiffnesses with hardware trials on stiffness-controlled TPU cubes. Grasp success, settled strain, carried strain, and damage are evaluated using calibrated motor-effort sensing and explicit slip and damage criteria.

  • Simulation setup: MuJoCo simulation uses the SO-ARM101 at 50 Hz, with cube side D∈{35, 40, 45} mm and stiffness k∈{2, 4, 6, 8, 10} × 10^3 N/m.Cube size and stiffness are sampled per trial; pad–fruit contact is calibrated to a Hookean F=kδ response with friction μ=5.0.
  • Simulation setup: The motor-effort model uses zero-mean Gaussian noise σ_I=0.003 A, with detection threshold ≈3.3σ, Npersist=3, a 10-tick baseline, and closing speed v=5 mm/s.These settings match the hardware detection configuration and operating point.
  • Experimental design: The proposed-method sweep tests six ε values across five stiffnesses with 50 seeds per cell, totaling 1500 trials; each fixed-force baseline contributes 250 trials.Each seed samples cube size and sensor-noise sequence, with greater trial-to-trial spread in soft cells.
  • Hardware setup: Hardware trials use SO-ARM101 with Feetech STS3215 servos at 50 Hz and 40 mm TPU cubes at 5 %, 10 %, and 15 % infill, labeled soft, medium, and stiff.The printed cubes keep geometry and material identical while varying stiffness through infill.
  • Evaluation criteria: Hardware success requires minimum top-of-lift motor-current proxy F̂ to exceed 10 load units (≈0.3 N at the jaw), while simulation uses an end-state slip check.Simulation success additionally requires the cube to rise at least half the gripper’s rise and settle its velocity.
  • Evaluation criteria: Damage is flagged when rhook=F/(k⋅D) exceeds 1.5 % for ≥5 consecutive simulation ticks; stop strain is the median over the final 10 ticks after a 20-tick hold.An analogous top-of-lift median gives carried strain, while single-tick contact-solver transients are rejected.

5. Results

The proposed strain-limited controller achieves damage-free grasping across medium-to-firm stiffnesses and outperforms fixed-force baselines, while very soft objects expose a trade-off between grasp reliability and damage. Detection latency imposes a strain floor that makes sufficiently small ε unattainable.

  • Detection floor: For ε≤1.3 %, settled strain remains 0.35–0.56 %, indicating contact-detection latency sets the stop; ε=0.3 % lies below this floor and produces unreliable grips.The controller cannot stop below the strain already spent detecting contact, so the requested bound is unattainable in that regime.
  • Hardware results: 40 % damage on the soft cube is lower than the 100 % damage of both fixed-force baselines, while grasp success matches them on medium and stiff cubes at roughly half the grip force.The proposed method applies approximately 62 load units versus 103 for fixed-low and 119 for fixed-high; medium and stiff cubes show 100 % grasp and 0 % damage except one 80 % cell.
  • Simulation results: 100 % grasp at 0 % damage occurs at k=4000 N/m for ε∈{0.7, 1.4, 2.0} %, with ≥98 % grasp at 0 % damage across ε∈[0.7, 2.0] %.Neither fixed-force baseline reaches the clean corner; the aggressive baseline applies ≈11 N and damages 100 % of grasps, while the gentle baseline achieves 68 % grasp.
  • Very-soft objects: At k=2000 N/m, every fixed-force threshold damages at least 70 % of the softest objects, whereas the proposed controller stays damage-free through ε≤1.4 %.Compression and damage increase monotonically with force threshold F_th; stop-on-contact is the least-compression fixed-force variant.
  • Simulation sweep: ≥98 % grasp at 0 % damage holds for every k≥4000 N/m and ε≥0.7 %, while fixed-force baselines do not match this sweep-wide result.The proposed method records 17 of 30 cells strictly at 100/0; the gentle baseline reaches the clean corner only for k≥6000 N/m.

6. Discussion

The controller provides certified, near-damage-free grasping across medium-to-firm stiffnesses using only encoder and motor-effort signals, while contact-detection latency creates a speed-dependent strain floor. Very soft objects remain a limitation, and stiffness adaptation is constrained by the excitation that damage-safe stopping suppresses.

  • Performance: ≥98 % grasp at 0 % damage holds for every k≥4000 N/m and ε∈[0.7, 2.0] % in calibrated-noise simulation, unlike fixed-force baselines.The method uses only motor-effort and encoder signals, without tactile or force/torque sensors.
  • Detection floor: 0.35–0.56 % measured strain persists across ε≤1.3 %, because robust contact detection consumes compression before the controller engages.The certified guarantee applies only for ε≥εfloor; at ε=0.3 %, stopping below the detection floor is unattainable by construction.
  • Limitations: 40/80/100 % damage occurs at ε=0.7/1.3/2.0 % on soft hardware cubes, while very soft objects require compression near the damage line for stable holding.Soft TPU cubes are proxies; biological produce has not yet been tested.
  • Detection floor: 0.10 % + 0.093 % ⋅ v [mm/s] describes the detection floor, making closing speed an operator-controlled throughput–gentleness knob alongside ε.Reducing only the persistence-window speed can achieve the same reduction without globally slowing the close.
  • Certified control and learning: F̂stop = εkminD converts a dimensionless strain specification and class-level priors into a certified, size-scaling stopping rule rather than a hand-picked force threshold.A live stiffness estimate is avoided because damage-safe stopping suppresses the excitation needed to identify stiffness; probing or learned estimation could trade exploration against damage.

7. Conclusion

The paper presents a sensorless compression-strain controller for damage-bounded grasping. Using encoder position and motor-effort signals, it stops gripper closure at a user-specified strain bound with a provably conservative compression estimate.

  • 7. Conclusion: The controller uses only encoder position and motor-effort signals, without tactile or force-torque sensors.This enables sensorless damage-bounded grasping.
  • 7. Conclusion: It estimates compression by dividing an effort-based force proxy by a known lower bound on object stiffness.The estimate supports a single proportional law for gripper closure.
  • 7. Conclusion: The gripper closes until estimated compression reaches a user-specified bound 𝜀.The strain bound directly defines the controller’s stopping condition.
  • 7. Conclusion: Because the operative stiffness is a lower bound, the stopping rule is provably conservative.The passage states that true compression stays below the conservative stop condition.
Loading 2608.23983v1…